Machine Learning in Traumatic Brain Injury: Dynamic Prediction from Continuous Physiological Monitoring and ICU Data.

Abstract

Background: Continuous physiological monitoring in neurocritical care generates large volumes of high-frequency time-series data that remain underutilized in routine traumatic brain injury (TBI) management. Machine learning (ML) approaches applied to these data have emerged as a promising strategy for dynamic prediction of secondary brain injury and individualized clinical decision support.

Methods: This narrative review summarizes current ML applications based on continuous physiological monitoring in TBI and related neurocritical care populations. A structured PubMed search supplemented by manual reference screening identified studies published through January 2026. Included studies applied ML methods to continuously acquired physiological signals such as intracranial pressure (ICP), arterial blood pressure (ABP), cerebral perfusion pressure (CPP), and related intensive care unit data.

Results: Thirty-four studies were identified, of which 26 met core inclusion criteria focused on continuous physiological monitoring. The most common applications were prediction of intracranial hypertension events, mortality, and functional outcome. Fewer studies addressed arterial hypotension prediction or unsupervised clustering approaches. Recent work demonstrates a transition from static admission-based prognostic models toward dynamic prediction using continuously updated physiological data. Deep learning architectures, including recurrent neural networks and transformers, were increasingly used for high-resolution time-series analysis. However, substantial heterogeneity exists in study design, preprocessing methods, prediction targets, and validation strategies. Most studies relied on retrospective observational data and internal validation, while prospective evaluation and workflow-integrated implementation remain limited.

Conclusion: ML approaches using continuous physiological monitoring show significant potential for dynamic prediction and decision support in TBI care. Future progress will depend on multicenter external validation, standardized methodological frameworks, prospective evaluation, and integration into real-time neurocritical care workflows.